The data comes from the CCF competition 爱奇艺剧本情感识别.
Aspect based emotion analysis is a rather complex task. In this competition, we employ a promote engining, where we construct sentences like "What's the emotion of XXX".
We only use RoBerta Chinese pre-trained language model, and add six regression layers for different emotions.<Happy, Angry, Sadness, Threat, Love, Afraid>.
We strongly recommend you don't modify the model itself unless you have enough confidence.
[CLS]+text+[SEP]+Aspect+[SEP]
text: we will combine the context of the same scene together
aspect: it's a question-like sentence.
We only train for one epoch, and get 0.7+ scores. If anyone consider using our code, you may get better result!